I'm currently writing a master thesis where I look at the predictive power of Google search volume in predicting the movement of the Norwegian Stock Market (Oslo Børs). I'm using Google search volume index (SVI) as a predictor (with other control variables) regressed against three measures: the stock return, the trading volume and the volatility. My sample consist 28 companies from the OBX index. The regression will be based on both cross-sectional and time-series data, thus it is necessary to arrange the data as panel data in order to analyse both types of data simultaneously. Panel data analysis is normally conducted with either fixed effect or random effect. And we run the Hausman test to see which of the two should be applied. So to my question, regardless what the Hausmen test tells, in my case which of those two effects should be applied? Or which one would make more sense to be applied?
-
$\begingroup$ Does Google search volume index indicate a direction? Ie you can get high volumes for, for example, Elon musk writing on twitter that he's taking his company private, or, for example. Elon musk having problems with the sec for securities fraud... Each of which had opposite effects. Are you looking to explain volatility? Or actually directional moves? $\endgroup$– willCommented Oct 6, 2018 at 0:44
1 Answer
I am actually unsure of your question.
Both the fixed effects and random effect models will depend on the assumptions that you make.
To keep it simple, in the Fixed effects model: all the individual differences are captured by differences in the intercept parameter.
In the Random effects model: all individual differences are captured by the intercept parameters but the individual differences are treated as random rather than fixed.
If random effects are present, it is preferred for several reasons, including,
1) The random effects estimator takes into account the random sampling process by which the data were obtained
2)The random effects estimator permits us to estimate the effects of variables that are individually time-invariant
3)The random effects estimator is a GLS estimation procedure, and the fixed effects estimator is a LS estimator.
So to answer your question, test for random effects and then see which model is preferred, but there is no clear you should use this model rather than that one.
-
1$\begingroup$ In the case above, the response is the stock market return ( well one of them ) and the predictor is the google search volume index. So, in that case, does one model, random effects or fixed effects, make more sense than the other ? I realize that there are tests available but I'm trying to understand when one is more appropriate. thanks. $\endgroup$ Commented Sep 6, 2018 at 1:26
-
$\begingroup$ I think the important thing to do is to test if random effects are present. If the random effects are present this makes it more appropriate. $\endgroup$ Commented Sep 6, 2018 at 11:33
-
$\begingroup$ To think about this more simply, if you are running a regression and there is heteroskedasictiy present, a sensible thing would be to adjust the variance covariance matrix to correct for this. If you are running a panel data model and there are random effects present you should use the random effects estimator. $\endgroup$ Commented Sep 6, 2018 at 11:34
-
$\begingroup$ Ok. I'll have to read more about panel data then. Thanks. $\endgroup$ Commented Sep 6, 2018 at 15:00
-
$\begingroup$ I can email you some lecture slides? $\endgroup$ Commented Sep 7, 2018 at 8:15